Mette Lise Lousdal and Henrik Støvring
While organised mammography screening programs were being gradually introduced across various countries, researchers could study the impact of screening on breast cancer mortality by comparing mortality in areas with and without screening. Now that screening has been fully implemented in most Western countries, researchers can only compare women who participate in screening with those who do not participate.
Women who do not participate in screening may seem to be a good choice as a comparison group, as they are not affected by screening. But the question is: can non-participants reflect how breast cancer mortality would have developed in women in general without the introduction of screening?
In our study recently published in the IJE, we employed variants of the inventive epidemiological tool of negative controls to assess the underlying breast cancer mortality in women who do not participate in screening in indirect ways.
When comparing non-participants with participants, we found that non-participants were more likely to die from causes other than breast cancer and also from external causes such as accidents, intentional self-harm and assaults. Intriguing! How can mammography screening in any way protect someone from dying in an accident? We therefore interpreted these protective associations as being an indication that non-participants generally had worse health than participants, including their underlying risk of dying from breast cancer.
We also found that going to the dentist on a regular basis apparently protected again death from breast cancer, irrespective of participation in mammography screening. As it is fairly inconceivable that a dentist can affect someone’s risk of dying from breast cancer, we again interpreted this as an indication that non-participants generally took less care of their health through preventive behaviour.
To ensure that all women included in our study were free from breast cancer at the beginning of the study, we only included women who had participated in their first-invitation round of screening and had a normal screening result. This meant that we were comparing women who declined their second screening invitation with women who did participate in the second round. We would expect the underlying mortality differences between all-time participants and non-participants to be even stronger than the differences we observed between second-round participants and non-participants.
Our findings clearly show that women who do not participate in screening are not representative of women in general. Therefore, women who do not participate in screening cannot readily be used as a comparison group in observational studies of screening.
Previous studies have used purported correction factors spanning from 0.64 to 1.36 to adjust for non-participants’ different risk of dying from breast cancer. However, these factors have varied widely across time and settings, which brings their applicability in specific settings into question. Further, once all women are invited for screening, such correction factors can no longer be estimated, and researchers therefore run the risk of relying on an ineffective correction without realising the risk of error. As a remedy, we suggest that negative control analyses, such as those done in our study, should be routinely undertaken to detect underlying differences in cancer mortality between groups.
Negative control analyses are relevant not only for studies of screening but, in general, for observational studies based on electronic health records.
Consider, for example, a Danish study that is based on registered prescription redemptions of oral contraception for women and the risk of venous thromboembolism (VTE). The researchers may be worried about the possibility of uncontrolled confounding, because women at elevated risk of VTE may choose other types of contraception. The difficulty, as in all such studies, is to come up with a suitable negative control exposure or outcome, which is available and satisfies the conditions. In this example, a suggestion for a negative control outcome might be to consider cycling accidents, as they may be inversely associated with overweight (a known but, in register-based studies, often unobserved risk factor for VTE) and are obviously unrelated to the choice of contraception.
We believe that researchers should become accustomed to thinking in this way, such that negative control analyses can become a standard precautionary measure to detect spurious associations.
Lousdal ML, Lash TL, Flanders WD, et al. Negative controls to detect uncontrolled confounding in observational studies of mammographic screening comparing participants and non-participants. Int J Epidemiol 2020; 25 March. doi: 10.1093/ije/dyaa029.